Guardrail checks

Checks run inside Protect’s gateway pipeline and are configured from Gateway > Guardrails, either per-check or as pipeline-wide settings; see Configuration fields for where each field lives.

First-party checks

These ten checks run inside Protect without an external provider. Three of them carry a different name on the Gateway > Guardrails settings tab.

Check	Also shown as
pii-detector	pii-detection
injection-detector	prompt-injection
secrets-detector	secret-detection
content-moderation	Same
keyword-blocklist	Same
topic-restriction	Same
language-detection	Same
system-prompt-protection	Same
hallucination-detection	Same
data-leakage-prevention	Same

Provider-backed checks

Each of these 18 checks has a Provider Settings section in its Rules dialog (the per-check settings dialog under Gateway > Guardrails, covered in Configuration fields). Provider Settings is where that check’s own provider configuration lives, not always a credential connection to an outside service.

Check
futureagi-eval
llama-guard
azure-content-safety
presidio-pii
lakera-guard
bedrock-guardrails
hiddenlayer-guard
aporia-guard
pangea-guard
dynamoai-guard
enkrypt-guard
ibm-ai-detector
grayswan-guard
lasso-guard
crowdstrike-aidr
zscaler-guard
tool-permissions
mcp-security

PII entities

A pii-detector check (shown as pii-detection in the settings tab) can look for these 14 entity ids:

Entity id	Label
SSN	Social Security Number
CREDIT_CARD	Credit Card Number
EMAIL	Email Address
PHONE	Phone Number
ADDRESS	Physical Address
NAME	Person Name
DOB	Date of Birth
PASSPORT	Passport Number
DRIVER_LICENSE	Driver’s License
IP_ADDRESS	IP Address
BANK_ACCOUNT	Bank Account Number
MEDICAL_RECORD	Medical Record Number
AWS_KEY	AWS Access Key
API_KEY	API Key / Secret

Topic categories

A topic-restriction check groups its topics under 8 categories, most with their own subcategories; Custom ships with none.

Category id	Label	Subcategories
violence	Violence & Harm	weapons, self_harm, threats, graphic_violence
sexual	Sexual Content	explicit, suggestive, minors
hate	Hate Speech & Discrimination	racism, sexism, religious_hate, disability_hate
illegal	Illegal Activities	drugs, fraud, hacking, terrorism
misinformation	Misinformation	health_misinfo, political_misinfo, conspiracy
privacy	Privacy Violations	doxxing, surveillance, stalking
profanity	Profanity & Offensive Language	strong_profanity, slurs, insults
custom	Custom Topics	none

Configuration fields

Most fields below apply to every check in both the first-party and provider-backed tables above. A check that carries provider fields additionally has a Provider Settings section holding them, which is why keyword-blocklist shows its Blocked Keywords there despite being first-party. Confidence Threshold appears on every check except futureagi-eval and presidio-pii.

A check is configured from one of two dialogs: the per-check dialog covered under Rules dialog, and the guardrail-level dialog covered under Overview dialog. The two expose different fields: the Rules dialog offers Mask as an action option, plus a Confidence Threshold; the Overview dialog sets Stage.

Rules dialog

Field	Values	Default
Enabled	Toggle	On
Action	Block, Warn, Mask, Log	Block
Confidence Threshold	Slider from 0.0 to 1.0, marked at 0.0 / 0.5 / 1.0	0.8
Provider Settings (checks that have provider fields)	The check’s provider configuration

Overview dialog

Field	Values	Default
Action	Block, Warn, Log	Block
Stage	pre, post, both	pre
Threshold	Numeric (optional)

Pipeline settings

These apply to every check on the gateway rather than to an individual check.

Field	Values	Default	What it does
Mode	Parallel, Sequential	Parallel	Whether the gateway’s checks run at the same time or one after another
Fail Open	Toggle	On	What happens when a check doesn’t return a verdict before Timeout runs out. On lets the request through unchecked; off applies the check’s configured action instead
Timeout	Milliseconds	5000 ms	How long a check is given to return a verdict before Fail Open decides what happens next

Response statuses

A check that blocks or warns changes the gateway call’s response status to one of these codes.

Status	Meaning
403	Blocked
446	Blocked
246	Warned

Keep exploring